Deciphering climate-induced displacement in Somalia: A remote sensing perspective

leave their homes with a delay after the drought starts. In addition, for those who opt to return home, it takes much longer, as it is likely that those migrating have decided that their previous living conditions were untenable. In conclusion, the impact of droughts on departure and return displacements is more com- plex than that of floods. However, reviewing the entire AOI period on drought and IDPs for all districts reveals that a district affected by a severe drought does not necessarily result in a high displacement of people. Discussion and concluding remarks In this study, we utilised remote sensing datasets to create various environmental indicators for Somalia including Agricultural Drought Index (ADI), flooding, and Land Cover (LC) changes. These were used to demonstrate the impact of extreme events on ground conditions in Somalia. Initially, the ADI methodology was developed around the cause-effect correlation found in agricultural droughts, where inadequate precipitation results in a soil moisture defi- cit, leading to decreased vegetation productivity. We generated a monthly time series of ADI, which corresponded to the different stages of the drought effects, categorised into three sever- ity levels: watch, warning, and alert. Drought monitoring precision could be improved by sup- plementing data input with in-situ measurements of higher spatial resolution or by consistently using remote sensing and model ensemble techniques. For flood mapping, a change detection and thresholding methodology was employed using,

Sentinel-1 imagery closely timed to flood incidents. Historical satellite data before the Sentinel era could potentially augment our analysis; however, our methodological constraints, along- side the discernible disparities in spatial and spectral resolutions between Landsat and Sentinel series, deterred the integration of Landsat data. Despite the absence of Sentinel data pre-2015, our study navigated these limitations by focusing on the robust capabilities and temporal alignment of available Sentinel-1 and Sentinel-2 data post-launch, ensuring consistency in our remote sensing analysis. The absence of pre-2015 Sentinel data and lack of driect ground truth data for validating the ADI and LC change map constitute significant limitations, reinforcing the need for future studies to explore complementary datasets where feasible, while maintain- ing analytical coherence. We developed a series of LC change maps to evaluate the repercussions of natural hazards, specifically drought and flood, on various land cover types, particularly those associated with agricultural productivity and human settlements. The LC methodology was constructed using the state-of-the-art U-Net technique, trained on the Sentinel-2 Level-2A dataset. The convolu- tional characteristic of U-Nets allows for the integration of information from surrounding pix- els into predictions, offering superior detection of large-scale structures within images, compared to the single-pixel detection of Random Forests. The U-Net training process was considerably expedited by collecting training data via poly-

gons, circumventing the laborious task of manually labelling pixels across an entire image. As a result, the final trained model outperformed conventional machine learning techniques typi- cally employed for land cover purposes, such as Random Forest. However, a notable limitation of LC change involves the inherent difficulty in disentangling the impact of drought from con- current events throughout the entire growing season. These concurrent events could include flooding and deliberate shifts in cultivation patterns, potentially complicating the interpreta- tion of the results. Our observations echo the findings of seminal works in this field [42–44], strengthening the validity of our assertions. Moreover, the nuances we’ve identified in the impact of concurrent events during the growing season, though briefly mentioned by [45, 46] PLOS ONE Deciphering Climate-induced Displacement in Somalia: A Remote Sensing Perspective PLOS ONE | https://doi.org/10.1371/journal.pone.0304202 August 9, 2024 22 / 27